Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS7421563 B2
Publication typeGrant
Application numberUS 11/208,882
Publication dateSep 2, 2008
Filing dateAug 23, 2005
Priority dateAug 23, 2005
Fee statusPaid
Also published asUS7818538, US8166278, US20070050596, US20080320271, US20110035566
Publication number11208882, 208882, US 7421563 B2, US 7421563B2, US-B2-7421563, US7421563 B2, US7421563B2
InventorsLaurence H. Cooke
Original AssigneeOc Applications Research Llc
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Hashing and serial decoding techniques
US 7421563 B2
Abstract
A technique for generating a list of all N-bit unsigned binary numbers by starting with an initial number less than some power of 2, successively multiplying the number by that power of 2 and adding the largest non-negative number less than that power of 2 such that the new number is not a duplicate of any of those already generated, and using the resulting lists to generate efficient hashing and serial decoding hardware and software.
Images(11)
Previous page
Next page
Claims(3)
1. A method of serially accessing all data locations of a contiguous linear storage array, using a single address, before repeating an access to a data location, the array being of a size equal to a power of two elements, the method comprising:
beginning with a current address of a location in said array, accessing data stored at the location corresponding to said current address;
generating a next address of a location within said array, said generating comprising doubling said current address and inserting a least-order bit of said data into a least-order bit position of a result of doubling the current address to obtain said next address;
setting said current address equal to said next address; and
iterating said accessing, said generating and said setting until all data locations have been accessed.
2. A method as in claim 1, wherein a portion of the data stored within the array is generated by sorting a backward-rotated list of least-order bits of an ordered list of numbers that cover the address space of said array, wherein said ordered list of numbers is generated by beginning with an initial value and successively choosing a power of 2 times said initial value plus an increment, modulo said size of said array, as a next number, wherein said increment is the largest number less than said power of 2 such that said next number has not already been chosen, and wherein said initial value is any non-negative number less than said power of 2, and wherein said next number replaces said initial value to generate another next number.
3. A method of defining connections in a serial shift register decoder for an array of size equal to a power of 2 elements, said decoder comprised of at least one circular shift register, the method comprising:
generating an ordered list of all numbers between zero and said size minus one, said generating comprising: beginning with an initial value and successively choosing a power of 2 times said initial value plus an increment, modulo said size of said decoder, as a next number, wherein said increment is the largest number less than said power of 2 such that said next number has not already been chosen, and wherein said initial value is any non-negative number less than said power of 2, and wherein said next number replaces said initial value to generate another next number;
defining address connections to bits of said at least one circular shift register according to low-order bits of the numbers of said ordered list; and
accessing said address connections to bits of said at least one circular shift register.
Description
FIELD OF THE INVENTION

The present invention pertains to decoding a serial stream of address data and traversing all entries of a Hash Table in a non-incremental fashion. Both the structure of the serial decoder and the data used to generate the non-incremental Hash Table addresses are derived from the results of an algorithm, which generates a list of 2N numbers in non-numerical order.

BACKGROUND OF THE INVENTION

Numerous examples of serial memory addressing exist including U.S. Pat. No. 4,667,313, granted May 19, 1984 to Pinkham et. al.; U.S. Pat. No. 5,452,255, granted Sep. 9, 1995 to Mine et al.; and U.S. Pat. No. 6,639,867, granted Oct. 28, 2003, but they all serially load a register with the address, and then use various combinations of the shift register data and traditional decode logic to select the specific word, as opposed to directly serially decoding the address.

Serial loading of addresses, as well as serial access of successive memory locations is becoming more important as the asynchronous high bandwidth nature of communications between chips and the cost of integrated circuit I/O coupled with the inherently slower on-chip clocking is resulting in a shift away from parallel synchronous input to high-speed Serialize/De-serialize (SERDES) input. Currently the external serial input is shifted into a register, which is driven in parallel into the chip when the last bit is captured. If the data is an address, it is decoded in parallel in order to access the memory. As a result there is the delay of the parallel decode and word access after the last external bit is available when accessing an on chip memory.

Alternatively, in high-speed systems, the address is broken into high- and low-order bits. Presumably the high-order bits are loaded first, and then the low-order bits are loaded. These low-order bits are powered up and drive successive levels of multiplexers to make the data available on the output shortly after the low-order bits are loaded into their register.

Reference is now made to FIG. 1. Many communications chips have high-speed serial interfaces, and some of the data being serially sent to such chips consists of address data. Typically this data is first converted from serial to parallel form in a SERDES 10. The address data is then transferred to the address registers 13,14 of an on-chip memory. In high performance memory designs, the high-order bits of the address data are loaded into a register 13 and decoded through a decoder 12, to drive one of the word lines in the memory core 11. The resulting data from that word line is then selected by loading the lower-order bits of the address data into a separate address register 14, which drives a series of multiplexers 15. By loading the high-order bits first, the data can be available on the inputs of the multiplexers 15 when the low-order address bits are loaded in their address register 14, limiting the access time to the delay from the address register 14 through the multiplexers 15. In simpler memories, both high- and low-order bits are loaded at the same time. In that case, the access time is the propagation time through the decode 12, memory core 11 and multiplexers 15. The translation from serial to parallel operation is required because this propagation time is typically much longer than a single serial clock cycle.

Using serial address data to access memories is a typical operation in high-speed serial communications routers and switches, which must be able to translate a large number of addresses into a limited number of port addresses or to indicate the address is not translated to a port in this network. Content-Addressable Memories (CAMs) have occasionally been used to do this type of operation. CAMs have the advantage of completing the search in one operation, but they require both a lot of time per access and a lot of transistors per memory element because they compare the inputs with all the words in their memories on each operation. An alternative is the hardware equivalent of a software technique called hashing.

Hashing searches a limited number of items that are stored in a small linear array when the address space for those items is much larger than the array itself by accessing the array with an index address that is both within the range of the array and is an arithmetic function of the item's full address. This implies that many items with different full addresses could map to the same index address. The problem then becomes storing the items efficiently enough in the array to maximize the utilization of the array while finding any given entry in the fewest number of memory accesses. Existing software hashing techniques suggest hashing tables contain a prime number of locations and the index for addressing into the hashing tables maps the large address space evenly into each of the locations. Since there could be multiple items in the table with the same index, only one of which is stored at the index address, it is also necessary to step through the memory in some fashion if the first item accessed is not the correct one. Many hashing algorithms suggest incrementing through the hashing table with another prime that is smaller than the size of the table. Ring theory shows that this will ultimately traverse through all elements in the table without repetition. One way to create an initial index and increment may be:
Index=Mod(original address, table size), and
Increment=some prime<table size

Now to determine the existence of an item in the table, perform the following:
Do until item=Table(Index) or Table(Index)=null
Index=Index+Increment
Loop

If the Table(Index) is null, then the item doesn't exist in the table; otherwise it does. In many applications, if it doesn't exist, it is added to the table. Clearly, the table begins with all null values and becomes more inefficient in determining if an item is in the table as the number of entries grows, since clashes exist with multiple items that map to the same hashing index.

Traditionally the Increment is a constant 1, which makes the next index easy to create, but often the data put into a hash table clusters about a few numbers, which makes the simple increment inefficient. Two solutions have been proposed in the past to eliminate this problem: either create an index that sufficiently separates otherwise adjacent entries, or create an increment that is both prime and varies with the entry in a way that is sufficiently different from the index to cause entries with the same initial index to have different increments.

The difficulty with this approach, when building hardware, is the need for a multiply or divide and the use of prime numbers, which are not easily addressable. Patents such as U.S. Pat. No. 6,804,768, granted Oct. 12, 2004 to Moyer, and U.S. Pat. No. 6,785,278, granted Aug. 31, 2004 to Calvignau, et al., describe ways to create indexes without doing multiplies or divides, but not the increment function.

The best solution would be to traverse a space that is N binary address bits, or 2N elements in size, ideally in a non-incremental fashion, and to do so without addition or multiplication. The problem then is to find a technique that is certain of traversing the entire size 2N space without stepping through successive addresses.

SUMMARY OF THE INVENTION

This disclosure describes a method for accessing all data within a contiguous linear array before repeating by generating each subsequent address using a function of the current address and the data at the current address within the array, where the array is a power of 2 in size, and the function is a single bit shift and low-order bit insertion.

More specifically, some portion of the data within the array is generated by sorting a backward-rotated list of the least-order bits of a list of numbers that cover the address space of the array, where the list is generated by beginning with the number zero and successively choosing between double the number, modulo the size of the array, and double the number plus one as the next number, where double the number is chosen if double the number plus one is already in the list.

The disclosure also describes a method for generating a ordered list of all numbers between −1 and some power of 2, by beginning with the number zero and successively choosing between double the number and double the number plus one as the next number, wherein double the said number is chosen if double the said number plus one is already in the list; and the least-order bit of said ordered list is used to define connections in a serial shift register decoder. This forms the mathematical basis of the above-described hashing technique.

The serial shift register decoder is comprised of a single address input; a multiplicity of word line outputs; and a multiplicity of shift register stages, where each of the shift register stages is connected to a previous and a next shift register stage such that the data shifts through the shift register stages in a circular fashion, the input of half of the shift register stages is the AND of the said address input and the previous shift register stage, and the input of the other half of the shift register stages is the AND of the inverse of the address input and the previous shift register stage, and one of the word line outputs is driven by each shift register stage.

In one mode, after serially applying all address bits to the serial decoder at one bit per clock cycle, one of the shift register stages of the serial decoder drives its word line to the opposite state of all the other word lines; and in another mode, after each clock cycle, the shift register stage that drives that one word line transfers its data to the next shift register stage.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will now be described in connection with the attached drawings, in which:

FIG. 1 is an abstract diagram containing a SERDES which drives a memory,

FIG. 2 is a diagram of a decoder,

FIG. 3 is a diagram of an embodiment of a serial decoder,

FIG. 4 is a diagram of one embodiment of the logic in each stage of the serial decoder,

FIG. 5 is a timing diagram of one embodiment of the serial decoder,

FIG. 6 is a diagram of another embodiment of the serial decoder,

FIG. 7 is a timing diagram of another embodiment of the serial decoder,

FIG. 8 is another embodiment of the serial decoder,

FIG. 9 is a diagram depicting one possible placement of the stages of a serial decoder,

FIG. 10 is another embodiment of a stage of a serial decoder,

FIG. 11 is a diagram of a serial decoder used to select between memory data lines,

FIG. 12 is a diagram of one embodiment of hash table hardware,

FIG. 13 is a diagram of a serial decoder with two parallel address inputs, and

FIG. 14 is another diagram of hash table hardware, which utilizes a multi-bit increment code.

DESCRIPTION OF VARIOUS EMBODIMENTS

Decoding is basically translating an address of N bits into 0s on 2N−1 lines and a logic level 1 on the specific line corresponding to the value of the N bits.

Reference is now made to FIG. 2. Traditionally, both polarities of the N input values 20 are generated by a pair of inverters 21 and drive 2N N-input AND gates 22, such that each AND gate is driven by a unique combination of one of the polarities of each of the original N values. For example, if the inputs 20 are A0,A1, . . . AN, then −A0, −A1, . . . −AN 23 are from the inputs and are all on an N-input AND gate, which may be used to select for address 0. Using 2-input NAND or NOR gates, this structure would take ˜N*2N gates.

A preferred embodiment of the present invention is a serial shift register-based decoder that takes fewer gates than the equivalent traditional decoder structure when decoding more than 4-bit addresses.

Reference is now made to FIG. 3. The shift register has 2N storage elements 31, and the data out from the last element 32 drives into the first element in a circular fashion. Each of the 2N word lines 30 is driven by a stage 31 of the shift register. A −S line 33, Clock line 34, and Enable line 35 drive into all stages of the shift register. An Address line 36 drives half of the stages in the shift register, and an inverter 37, which is driven by the Address line 36, drives the other half.

Reference is now made to FIG. 4, which shows the detailed logic of one embodiment of a single stage of the shift register decoder. The value in each stage of the shift register is driven out the Do line 40, which is connected to the Di line 41 of the next stage of the register. Prior to loading the first address bit, all storage elements are set when the −S line 44, goes low. Subsequently, loading each address bit corresponds with clocking the shift register one location. The previous stage's value on the Di line 41 is gated with the current address bit 42 (A) or its inverse (−A), and is captured in the current stage when the clock 43 is high. The word lines 45 are enabled when an enable signal 46 goes high after the last address bit is loaded. This embodiment of a stage of the shift register decoder takes 4.5 gates, so in this case a decoder would take 4.5*2N+1 gates. In other words, for N>4, this technique takes less logic than the traditional decoder.

Furthermore, a traditional N-bit decoder built out of 2-input NAND gates has a delay path from its address register that includes driving 2N−1 loads, and an Inverter that drives 2N−1 loads, where each load further drives Log2(N) levels of NAND gates. By contrast, the serial decoder's address input drives 2N−1 loads and an Inverter, which drives 2N−1 loads, where each load further drives a single 2-input NAND gate, before capturing the result in the shift register, and only drives one 2-input NAND gate after clocking the shift register. In other words, the delay through a traditional decoder, after capturing all its address bits, is always greater than the delay through the serial decoder presented, after it has captured its last address bit.

The key to a properly functioning shift register-based serial decoder is the sign of the Address bit inputted into each stage of the shift register. Each successive address bit must be entered using the same structure, while the resulting decoded address shifts one location around the circular shift register; but in a traditional decoder, each address bit stays in a constant location and drives different AND gates with different polarities, so a simple upper half—lower half decode for each bit doesn't work on a shift register-based serial decoder.

What is needed is a way to distribute all possible combinations of +/−Ax (i.e., the address bits and their inverses) such that each location has a unique combination of one or the other polarity of each of the A0 to AN bits, where each term is a successive shifted serial combination of each address bit.

It is mathematically known that a set of all permutations of +/−Ax for any number of Ax bits has 2N entries for N Ax bits, and many orderings of such a set are possible, but a specific ordering must be found to have the characteristics of a decoder. One way to find such an ordering is to create a simulation, and to select the polarity of each stage so that after N shifts, all permutations of +/−Ax exist in the shift register. Table 1 below contains three small examples of such a simulation. The +/−column indicates the polarity of the address bit ANDed at each stage of the shift register. The bits a, b, c and d correspond to the serial address bits A0,A1,A2, and A3 and the upper case corresponds to the negative polarity (i.e., inverse).

TABLE 1

Each column to the right of the gray column is a shift stage, where the bit loaded on that cycle is at the top of the column. In these three examples, the polarities were selected such that the N-bit address is translated into 2N possible word lines, where after N shifts all combinations of both polarities of each address bit exist, but not in normal numerical order. Also notice that the simulation is circular, as is the logic in FIG. 2, because the first location contains the negative polarity variables for all but the last inputted address bit, which it got from shifting the last bit in the shift register back to the first bit.

Still, in order for the decode technique to be useful, it is necessary to generate such a structure for 2N shift elements, where N is larger than 4, but in order to create such a generator it is necessary to better define the problem.

This type of decoder generates an ordered list of the 2N numbers, but it does so by successively combining the bits into stages of a circular shift register. If we transform the symbolic simulations in Table 1 into numbers, assuming 1s were shifted in, each entry in the last column of each simulation would correspond to a binary representation of a number from 0 to 2N−1. Table 2 shows the results of such a conversion from Table 1.

TABLE 2

Closer inspection of these results shows that each shift corresponds to doubling the previous stage's number and adding a 1 or 0 depending on the polarity of the stage of the shift register, and while there are duplications of numbers prior to N shifts, the results after N shifts are such that each number in each location is unique.

The key to generating such a list of polarities should be to create a 2N list of numbers by setting zero as the first element in the list and successively doubling, modulo 2N, and adding one to the previous number in the list, checking to see if the new number already exists in the list, and decrementing it if it does. Table 3 shows the results of such a simulation for the three cases in Tables 1 and 2.

TABLE 3

The first column in each example contains the number of the shift stage. The second column contains an X where a 1 was added to the number in the previous stage, and the third column contains the number for that stage of the shift register. As can be seen in Table 3, all N-bit numbers are represented in the 2N locations in each array, and though not in normal numerical order, correspond to the numbers in Table 2.

The code for a generation algorithm, which creates a list of 2N elements corresponding to the 2N stages of the shift register and generates a polarity array, where each element contains either a 1 for the positive polarity of Ax or −1 for the negative polarity of Ax, from that list is as follows:

List(1)=1;
For I=2 to 2N;
 Value=List(I−1)*2+1 Modulo 2N;
 For J=1 to I−1;
  If Value = List(J) then
   Value = List(I−1)*2 Modulo 2N;
   J=I
  end if;
 Next J;
 List(I) = Value
Next I;
For I=1 to 2N;
 Polarity(I)=[List(I) Modulo 2 ]*2−1
Next I;

Reference is now made to FIG. 5, a timing diagram of the decode structure shown in FIG. 3. The successive Address values (A) 50 are changed prior to the rising edge of the Clock (C) 51, which captures them into the shift register. After the last Address bit is loaded, the enable line (E) is brought high 52, which enables the Word lines (L) 53 to select the word being read or written, in which, when requiring more than one clock cycle, the clock is turned off 54 during the actual memory operation. Prior to loading the next address, the Set line (−S) must be pulled low 55 to reset the entire shift register.

Reference is now made to FIG. 6. After a specific address is loaded into the serial shift decoder, the resulting address may be shifted in a circular fashion throughout the entire decoder, by setting the Address data on all stages of the shift register high. This can be done by substituting the inverter 37 in FIG. 3 with a NAND gate 66 in FIG. 6. When the −I input 61 is set low and the Address input 62 is set high, both polarities of the Address input to each stage of the shift register are set high. Thus, after applying a positive pulse on the clock line 63, the single stage that is set high, which corresponds to the original selected address, is shifted to the next stage in the shift register.

Reference is now made to FIG. 7, which shows the timing diagram of the selection of a specific address followed by incrementally addressing a number of words in the memory. In this case, following the initial memory operation the Address input (A) and the increment control (−I) are set high 70 and low 71 respectively. Thereafter, each successive Clock (C) 72 causes the next address in the shift register to be selected. A simulation of the logic shown in FIG. 4 and extended in FIG. 6, is shown in Table 4 below.

TABLE 4
Real Address:
1 3 7 15 14 13 11 6 12 9 2 5 10 4 8 0
C# −I A −S 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Comments
0 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Initialize
1 1 1 1 1 1 1 1 0 1 1 0 0 1 0 1 0 0 0 0 Decode
2 1 1 1 0 1 1 1 0 0 1 0 0 0 0 0 0 0 0 0 Addr = 14
3 1 1 1 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0
4 1 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
5 1 1 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 Shift
6 0 1 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 14->13->
7 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 11->6
8 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Initialize
9 1 0 1 0 0 0 0 1 0 0 1 1 0 1 0 1 1 1 1 Decode
10 1 1 1 1 0 0 0 0 1 0 0 0 1 0 1 0 0 0 0 Addr = 5
11 1 0 1 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0
12 1 1 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

The columns labeled 0 through 15, show the logical state of each of the 16 shift stages. The clock is implicitly one cycle per row, as labeled by the number of clocks (C#). The Enable line, not present in the simulation, merely transfers the logical state of the 16 shift stages onto the Word lines, which would normally occur on cycles 4,5,6,7 and 12 of a simulation like the one shown in Table 4.

Note that as each bit of the 4-bit addresses is entered, the number of shift stages that are still set to a one level is reduced to half of the number on the previous cycle. Eight stages are set to one on cycles 1 and 9, four are set on cycles 2 and 10, two are set on cycles 3 and 11, with the remaining address set on cycles 4 and 12. Also note that on each cycle, the remaining stages set to a one level are rotated to the right one stage.

This serial decoder is designed for high-speed operation. The shift stages naturally want to be adjacent to each other, and by interleaving them in a fashion shown in FIG. 8, there is very little wire and only one gate delay between each stage of the decoder. While the clock is a regular signal with well-defined techniques for being distributed with very little skew, and the other control signals can also be tuned to reach all the stages of the shift register at their required times, the polarities of the address signal are distributed irregularly throughout the serial decoder, and may take more than one cycle to reach all the stages of the shift register.

Reference is now made to FIG. 8. To allow for the best possible performance, the Address line fan-out structure may be pipelined with the addition of 2K additional latches or flip-flops 80, which divides up the distribution of the Address signal into K sections, sufficiently small enough portions of the decoder to be processed in a single cycle. As a result an N-bit address will take N+1 cycles to decode, but at a shift speed that is comparable to the serial input data speeds from a SERDES.

Reference is again made to FIG. 1. Still this high performance memory has two address registers 13,14 for the upper- and lower-order address bits, and while a faster way to decode the upper bits has been shown, if the lower bits are loaded last and the delay path from the lower-order bit register 14 goes through the multiplexers 15, not a decoder. By replacing the multiplexers with tristate buffers, a faster serial decode solution can be implemented.

Reference is now made to FIG. 11. In this case, each of the word lines drives two multiplexers 110, which are controlled by a Last signal 111 that selects the prior word line on the penultimate cycle and the current word line on the last cycle of address loading. The outputs of the multiplexers 110 control tristate buffers 112, which in turn select the data line value from the memory core 113 to be driven onto the output 114. This serial decoder should have its enable signal set high at the end of the next-to-last address loading, which will enable two addresses, the word line prior to the actual address and the word line prior to the address with the opposite value for the address' least order bit, which is loaded on the last cycle, as shown in the simulation results in Table 4 above. When the last address bit is loaded, only the word line that corresponds to the final address is enabled. As such, with Last set low on the next-to-last cycle, two tristate buffers are enabled, one of which is still enabled on the last cycle, when the Last signal is set high. As a result, the data on the output 114 is either at mid voltage (of the tristate buffer) or the correct level one cycle ahead of loading the final address bit.

Now the delay of the traditional bit select and serial decode can be compared. The worst case delay path for the low-order address bits of the memory in FIG. 1 is 2N loads on the first register output driving Log2(N) levels of multiplexers to the data output. By contrast, the serial shift decoder's address is latched into the shift register after decoding, so the delay path from the latch is merely a 2-input NAND gate that drives a multiplexer, which turns off the tristate buffer that is not addressed, because in the prior cycle the two possible tristates, one of which becomes the correctly addressed tristate were both turned on. Not only is the serial decode's delay clearly shorter, in the best case, when both bits selected on the prior cycle are at the same value, there is no delay because the data was already there.

Reference is again made to FIG. 4, which shows one implementation of a stage of the serial decoder that requires a short clock pulse, because it is a single latch. A static flip-flop version of this logic would require at least 2 additional gates, but other smaller implementations are possible, such as a dynamic flip-flop as shown in FIG. 10, which only takes 3 gates. In other words, for low gate count implementations, serial decoders can require fewer gates than traditional decoders for anything above an 8:1 decode.

In yet another embodiment, the serial decoder, when implemented on a memory with its shift address option as shown in FIGS. 6 and 7, can serve as a hashing table.

In yet another embodiment of the current invention, the technique used to define the polarity of the address bits in the serial decoder may be used in conjunction with a normal memory to create a fast hardware or software hashing algorithm.

In general, the best hashing algorithm would be to traverse a space that is N binary address bits, or 2N elements, in size, ideally in a non-incremental fashion, and to do so without additions or multiplications. The problem is to find a technique that is certain of traversing the entire size 2N space without incrementing by 1 through all the addresses, and ideally without using an add or multiply operation.

The basis behind the serial decoder is a double and increment, which can be accomplished by a shift operation and least-order bit insertion. The examples in Table 3 above show full non-incremental traversals of the entire 2N address space without using a multiply or add to generate the successive numbers. So by taking the 16-bit section of Table 3, as shown in columns A, B and C in Table 5 below, rotating the Xs to one location prior to their current locations in column B, as shown in column D, changing the Xs to 1s and blanks to 0s, and placing them in the locations specified by the originally generated numbers shown in column C, one obtains a list of 1s and 0s as shown in column F, for inserting into the least-order bit to generate the next hash address after shifting the current address. For example, the first four rows of column B are marked with Xs. These are moved to the last row and the first three rows of column D. Now using the corresponding values in column C, as addresses, 1s are placed in locations 0, 1, 3 and 7 in column F, where column E shows the addresses of the table in column F. Similarly, the other four Xs from the top to the bottom of column D cause 1s to be placed in locations 14, 13, 12 and 2 in column F, with all other rows of column F being set to 0s. In this manner, the columns D through G in Table 5 show the transformation, necessary to generate a hash table with the last column, G, showing the computed next address by shifting the current address in column E and inserting the 0 or 1 increment from column F.

TABLE 5

It should be noted here that by successively going to the location specified in the right-most next column, all 16 locations in the table are accessed before repeating. Now, using the following procedure, the list of numbers can be generated:

Next Address(0)=0;
For I=1 to 2N−1;
 Next Address(I) = Next Address(I−1)*2+1 modulo 2N;
 For J=1 to I−1;
  If Next Address(I)=Next Address(J) then
   Next Address(I)=Next Address(I)−1
   J=I
  end if;
 Next J;
Next I;

This algorithm searches through the list of numbers already generated for one plus double the last number. If the number already exists, then double the number is used. It generates a non-repeating ordered list of numbers from 0 to 2N−1. The proof is not derived here, but the results of this algorithm can be easily checked by modifying it to fail when any of the even numbers are not unique.

Now the list of numbers can be used to create a set of increments that can be used in the following hashing algorithm:
Do until item=Table(Index) or Table(Index)=null
Index=Inc(Index)+Index*2
Loop

To create the Increment table, merely do the following:
Do for I=0 to 2^N−1;
Inc(Next Address(I))=Next Address(I+1 modulo 2N) modulo 2;
Next I

The Increment table could, for example, be one bit in the hash table.

Reference is now made to FIG. 12. The Increment array is loaded into the left-most bit of the Memory 120. An index 121, which is between 0 and 2N−1, the address space of the memory, is generated from the original item 122 and loaded into the Address shifter 123. In this example, Compare 124 compares the item 122 with the lower-order bits from the Index location in the Memory 120. If the compare is not equal 126, then the Index is shifted, and the Increment value 128 is inserted in the lowest-order bit of the Address shifter 123. This process continues until either a null is detected or the compare is equal. If a null occurs, the memory may be updated, by using the null signal 125 to control the read/write of the memory and putting the item with the current Increment value on the Data In port 127 of the memory; otherwise, the current address is in the address shifter 123.

It should be noted here, that the serial decoder automatically does this operation when moving from decode to serial shift mode.

This technique may also be implemented in software, and as shown in FIG. 12, the Increment bits are placed in the highest-order bit of the memory words. Given a reasonable function to create the Index from the Item, such as the exclusive-or of the Item's high-order bits with its low-order bits, this hashing technique may be performed on a processor with simple Boolean functionality, since no multiplies or adds are needed. This is very useful in the case where a dedicated embedded processor may be used to perform the hashing function, because a much smaller processor would meet the computational requirements. Below is an example of abstract assembly code that could be used in such a processor. In this example, the 2N Hashing Table is located on a 2N memory address boundary to eliminate the need for adding the base address of the table to the index.

Next: Load R1 From Memory(Value) /* R1 contains the Index
Load R2 with R1 /* R2 contains the Item
Shift Right R1 N bits /* R3 contains the Result
XOR R1 with R2 /* R4 contains the Increment
Find: AND R1 with 2N−1 /* Memory(Item) is a fixed
address
OR R1 with Table Address /* 2N−1, and Sign bit are
constants
Load R3 From Memory(R1)
Load R4 with R3
AND R4 with Sign bit  /* leave Increment bit only
XOR R3 with R4 /* clear Increment bit from
Result
Compare R2 with R3
Branch on equal to Found
Compare R3 with Null
Branch on equal to Insert
OR R1 with R4 /* put Increment into Index
Rotate Left R1 1 bit
Branch to Find /* Item was not found, continue
Insert: Branch on -Add to Next /* -Add is a predefined branch code
OR R2 with R4 /* Item was not found add it to table
Store R2 in Memory(R1)
Branch to Next
Found: /* Item was found in the Table

Clearly, the assembly code example above presumes the Hash Table entries were initialized with the Null value concatenated with the Increment value. Since the Increment bit is the sign bit, this code assumes the Items and Results are always positive (including the Null value), because it clears the Increment bit from the Result before comparing Item to it. It also assumes that the Table Address is positive because it puts the Increment bit into the high-order bit of the Index before rotating. It is further understood that a person skilled in the art may further optimize the code for any specific processor, and may modify it to handle Table insertions and/or subsequent searches in a different manner than presented above. Still, the above example does illustrate the technique's ability to traverse the Hash Table in a non-incremental way, without the use of Add or Multiply instructions.

In yet another embodiment of the present invention, serial decoders may consist of multiple circular shift register strings. The techniques described above may apply to any number of bits shifted in parallel into the decoder.

Reference is now made to FIG. 13. In this case, there are two address bits 130 loaded in parallel into two 16-bit circular shift registers, where each stage 132 is composed of the logic shown in FIG. 4. In FIG. 13, the outputs from any specific stage number, say 15, in both registers 132 and 133, drive the same AND gate 134, which in turn drives a word line 135. The polarity of each respective address bit is shown by the + or − in each of the stages. In this case 4 bits of address are loaded in 2 clock cycles, two bits per cycle. By generalizing the polarity generation for any M bits per cycle and any N cycles, the size of the decode is 2M*N and the polarities for each shift register and stage can be generated as follows:

List(1)=1;
For I=2 to 2N*M;
 Inc = 2M−1;
 Value=List(I−1)*2M+Inc Modulo 2N*M;
 For J=1 to I−1;
  If Value = List(J) then
   Inc=Inc−1;
   Value = List(I−1)*2M+Inc Modulo 2N*M;
   if Inc>0 then
    J=I;
   else
    J=1;
   end if;
  end if;
 Next J;
 List(I) = Value
Next I;
For I=1 to 2N*M;
 Value = List(I) Modulo 2M
 Do J=1 to M;
  Polarity(I,J)=[Value Modulo 2 ]*2−1
  Value = integer of Value/2
 Next J;
Next I;

Like the ordered list of numbers shown in Table 3, the next value in the list is determined by shifting the current value and adding the increment, but in this case the number of bits shifted is greater than one (if M>1) and the increment could be as large as 2M−1, so the algorithm above tries to use the current value*2M+ the largest increment (2M−1) first. If it finds that this value was already in the list, it then tries one less, and so on, until the current value*2M is selected as the next value. It should also be noted here that while all the examples in Tables 2,3 and 5 show the ordered lists starting with 1, actually, a complete ordered list may be generated when starting with any number from 0 to 2M−1. As with the single bit case, the technique always covers all values from 0 to 2MN−1. Also, the array Polarity contains 1 for the positive of the address bit and −1 for the negative of the address bit, for each of the 2N*M bits, but in this case for all the circular shift registers. The first dimension of the Polarity array is the stage number, and the second is the shift register number. For the decoder in FIG. 13, N=2 and M=2, hence 2N*M=16. The value of Inc can be either 0,1,2 or 3, a two-bit quantity that is broken into an upper and lower bit ring, where the upper bit ring is driven by A1 and the lower bit ring is driven by A0. The ordered list of numbers, increments and polarities, generated for the decoder in FIG. 13 by the code above, can be seen in Table 6 below, along with a simulation of the decoder:

TABLE 6
Increment 1 3 3 2 3 1 2 2 1 1 0 3 0 2 0 0
Polarity1: + + + + + + + +
Polarity0: + + + + + + + +
Real Addr; 1 7 15 14 11 13 6 10 9 5 4 3 12 2 8 0
C# Addr 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 A1 x 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
0 A0 x 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
0 L out: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 A1 1 0 1 1 1 1 0 1 1 0 0 0 1 0 1 0 0
1 A0 0 0 0 0 1 0 0 1 1 0 0 1 0 1 1 1 1
1 L out: 0 0 0 1 0 0 1 1 0 0 0 0 0 1 0 0
2 A1 0 0 0 0 0 0 1 0 0 1 0 0 0 1 0 1 0
2 A0 1 1 0 0 0 1 0 0 0 1 0 0 1 0 0 0 0
2 L out: 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0

The first row of Table 6 shows the Inc values for the 16 word lines. This translates into the polarity lists 0 and 1 applied to the rings in the example, the next two rows in the table. The fourth row in the table is the real decoded address, with the numerical order of the stages shown in the fifth row. Like the simulation shown in Table 4, the rings begin an address decode by being set to all 1s, but in this case the 4 to 16 decode shows the 4-bit address being inputted in two consecutive cycles of two bits each, where A1 on the first cycle is the highest order bit of the address, and A0 on the second cycle is the lowest order bit of the address. This example shows a decode of the number 9, which is located in stage 8.

Clearly, this approach takes almost twice the gates of a single-bit decode, but can decode two bits at a time, and the whole address in half as many cycles as a single-bit serial decoder. Furthermore this technique may be generalized as described above to generate a decode comprised of any M circular shift registers each with 2M*N stages, which can decode an M*N bit address in N clock cycles at M bits per clock cycle. For example, three 64-stage circular shift registers may be generated, which decodes a 6-bit address in two clock cycles (M=3, N=2, and 2M*N=64).

In yet another embodiment of the present invention, the hashing increment may also be more than one bit. A two-bit increment for a hash code increment can be generated from the Inc values in Table 6 in a similar fashion to generating the single-bit increment values. By rotating the values in the left direction one slot and sorting the resulting numbers by the real decode values, the following Table 7 is generated:

TABLE 7
Inc 1 3 3 2 3 1 2 2 1 1 0 3 0 2 0 0
Addresses 1 7 15 14 11 13 6 10 9 5 4 3 12 2 8 0
Rotated 3 3 2 3 1 2 2 1 1 0 3 0 2 0 0 1
Sorted
Addresses 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Increments 1 3 0 0 3 0 2 3 0 1 1 1 2 2 3 2

The Inc array shown in the first row of Table 6 is reproduced as the top row of Table 7, along with the real addresses from the fourth row of Table 6 in the second row of Table 7. When the Inc array values are shifted left one, as shown in the third row of Table 7, and both the real addresses and the increments are sorted by the real addresses, and the resulting increment order can be found in the bottom row of the table.

Reference is now made to FIG. 14. The resulting two-bit increment values in the bottom row of Table 7 can be loaded into the Memory 144, such that when the next address is needed, the current address 142 that was loaded in the decoder 143 is shifted two bits, and the two increment values 141 are inserted into the two lower-order bits. In this way the entire memory can be traversed in the same manner as with the single-bit case. Of course, similar changes to the software hashing are also contemplated to adapt the software hashing technique shown above to handle more than one bit of increment data.

It is also contemplated that the generated polarity lists may also be used in reverse order, or in any other order that meets the requirements to create serial decoders and hashing structures such as described above.

It is further contemplated that other serial functions may be generated comprising one or more features of the above serial decoder, or hashing table, such as serial comparators, serial selectors, serially addressed switches, serially addressable polling structures, or other similar structures.

It will be appreciated by persons skilled in the art that the present invention is not limited by what has been particularly shown and described hereinabove. Rather the scope of the present invention includes both combinations and sub-combinations of various features described hereinabove as well as modifications and variations which would occur to persons skilled in the art upon reading the foregoing description and which are not in the prior art.

Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US4667313Jan 22, 1985May 19, 1987Texas Instruments IncorporatedSerially accessed semiconductor memory with tapped shift register
US5452255May 18, 1994Sep 19, 1995Kabushiki Kaisha ToshibaSemiconductor memory device
US6639867Mar 26, 2002Oct 28, 2003Hynix Semiconductor Inc.Decoder circuit in a semiconductor memory device
US6785278Dec 10, 1998Aug 31, 2004International Business Machines CorporationMethods, systems and computer program products for hashing address values
US6804768Apr 15, 2002Oct 12, 2004Hewlett-Packard Development Company, L.P.Programmable microprocessor cache index hashing function
US6907439 *Mar 26, 2002Jun 14, 2005Lattice Semiconductor CorporationFFT address generation method and apparatus
Classifications
U.S. Classification711/219, 711/217, 711/218, 712/E09.039, 711/216
International ClassificationG06F12/00
Cooperative ClassificationG06F9/345
European ClassificationG06F9/345
Legal Events
DateCodeEventDescription
Sep 23, 2011FPAYFee payment
Year of fee payment: 4
Dec 7, 2010ASAssignment
Free format text: MERGER;ASSIGNOR:OC APPLICATIONS RESEARCH LLC;REEL/FRAME:025467/0063
Effective date: 20101207
Owner name: INTELLECTUAL VENTURES I LLC, DELAWARE
Apr 4, 2008ASAssignment
Owner name: OC APPLICATIONS RESEARCH LLC, DELAWARE
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ON-CHIP TECHNOLOGIES, INC.;REEL/FRAME:020753/0337
Effective date: 20080313
Owner name: OC APPLICATIONS RESEARCH LLC,DELAWARE
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ON-CHIP TECHNOLOGIES, INC.;US-ASSIGNMENT DATABASE UPDATED:20100323;REEL/FRAME:20753/337
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ON-CHIP TECHNOLOGIES, INC.;US-ASSIGNMENT DATABASE UPDATED:20100330;REEL/FRAME:20753/337
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ON-CHIP TECHNOLOGIES, INC.;REEL/FRAME:20753/337
Aug 23, 2005ASAssignment
Owner name: ON-CHIP TECHNOLOGIES, INC., CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:COOKE, LAURENCE H.;REEL/FRAME:016915/0403
Effective date: 20050822